# Decision: validate all LangChain batch identities before inserting The adapter rejected IDs already in the store but accepted repeated IDs within one add_texts call. Two native rows could share one LangChain ID; deleting that ID removed only the last row and left the earlier row without a forward mapping. Seed a seen-ID set from existing identities and extend it while validating the batch. Reject collisions before embedding or native mutation. Distinct explicit IDs and generated IDs keep their existing behavior. The regression exercises public add/search/delete calls backed by the published native extension and checks native row counts. Existing-ID rejection also preserves the original row. No persistent mapping format or LangChain upsert contract is introduced.